Embedding clinical event sequences enhances infection risk prediction in asplenic patients

Abstract

Asplenic patients are predisposed to severe infections and other harmful health outcomes due to their compromised immune function. Accurately predicting these risk events is crucial to facilitate timely and personalised clinical interventions. This study aims to evaluate embedding-based strategies for representing temporal clinical data obtained from Electronic Health Records. The ultimate goal is to predict the onset of infections in asplenic patients from a real-world Italian Network of Asplenia, which is a nationwide hospital multicentre database collecting EHRs from a cohort of 1789 asplenic patients, reflecting the heterogeneity and challenges of routine clinical data, including missing values, inconsistencies, and irregular sampling. The study leverages several state-of-the-art deep learning models and statistical methods to produce embedding-based representations that capture temporal dependencies within sequences of clinical events, including RNN-based, attention-based, transformer-based, and representation learning approaches, which are systematically compared within a unified framework. As a benchmark method for our experiments, we adopted a customary representation of clinical trajectories that is an occurrence vector encoding the presence or absence of events, whilst ignoring any temporal order. We assessed all approaches using a gradient boosting classifier (LGBM), explicitly decoupling representation learning from downstream prediction to enable a controlled and fair comparison, and systematically compared their predictive performance. The comparison highlights the advantages and disadvantages of using temporal (time-based) data representations for clinical risk prediction, and is complemented by interpretability analyses to support clinical insight, while offering interesting insights into their effectiveness in precision medicine. By combining a real-world dataset, a comprehensive comparative design, and a reusable evaluation pipeline, this work provides practical guidance for applying embedding-based approaches in clinical risk prediction.